Deep learning-based virtual noncontrast CT for volumetric modulated arc therapy planning: Comparison with a dual-energy CT-based approach.

PURPOSE: The aim of this study was to develop a deep learning (DL) method for generating virtual noncontrast (VNC) computed tomography (CT) images from contrast-enhanced (CE) CT images (VNCDL ) and to evaluate its performance in dose calculations for head and neck radiotherapy in comparison with VNC images derived from a dual-energy CT (DECT) scanner (VNCDECT ). METHODS: This retrospective study included data for 61 patients who underwent head and neck radiotherapy. All planning CT images were obtained with a single-source DECT scanner (80 and 140 kVp) with rapid kVp switching. The DL-based method used a pair of virtual monochromatic images (VMIs) at 70 keV with and without contrast materials. VMIs without contrast materials were used as reference true noncontrast (TNC) images. Deformable image registration was used between the TNC and CE images. We used the data of 45 patients, chosen randomly, for training (7922 paired images), and data from the other 16 patients as test data. We generated the VNCDL images with a densely connected convolutional network. As the VNCDECT images, we used VMIs with the iodine signal suppressed, reconstructed from the CE images of the 16 test patients. The CT numbers of the tumor, common carotid artery, internal jugular vein, muscle, fat, bone marrow, cortical bone, and mandible of each VNC image were compared with those of the TNC image. The dose of the reference TNC plan was recalculated using the CE, VNCDL , and VNCDECT images. Difference maps of the dose distributions and dose-volume histograms were evaluated. RESULTS: The mean prediction time for the VNCDL images was 3.4 s per patient, and the mean number of slices was 204. The absolute differences in CT numbers of the VNCDL images were significantly smaller than those of the VNCDECT images for the bone marrow (8.0 ± 6.5 vs 175.1 ± 40.9 HU; P < 0.001) and mandible (20.3 ± 19.3 vs 106.2 ± 80.5 HU; P = 0.002). The DL-based model provided the dose distribution most similar to that of the TNC plan. With the VNCDECT plans, dose errors >1.0% were observed in bone regions. The dose-volume histogram analysis showed that the VNCDL plans yielded the smallest errors for the primary target, although dose differences were <1.0% for all the approaches. For the maximum dose to the mandible, the mean ± SD errors for the CE, VNCDL , and VNCDECT plans were -0.13% ± 0.23% (range: -0.46% to 0.31%; P = 0.037), -0.01% ± 0.22% (range: -0.40% to 0.36%; P = 1.0), and 0.53% ± 0.47% (range: -0.21% to 1.41%; P < 0.001), respectively. CONCLUSIONS: In this study, we developed a method based on DL that can rapidly generate VNC images from CE images without a DECT scanner. Compared with the DECT approach, the DL-based method improved the prediction accuracy of CT numbers in bone regions. Consequently, there was greater agreement between the VNCDL and TNC plan dose distributions than with the CE and VNCDECT plans, achieved by suppressing the contrast material signals while retaining the CT numbers of bone structures.